Itâ€™s far more likely that robots would inadvertently harm or frustrate humans while carrying out our orders than they would rise up against us.
When we look at the rise of artificial intelligence, itâ€™s easy to get carried away with dystopian visions of sentient machines that rebel against their human creators. Fictional baddies such as the Terminatorâ€™s Skynet or Hal from 2001: A Space Odyssey have a lot to answer for.
However, the real risk posed by AI â€“ at least in the near term â€“ is much more insidious. Itâ€™s far more likely that robots would inadvertently harm or frustrate humans while carrying out our orders than they would become conscious and rise up against us. In recognition of this, the University of California, Berkeley has this week launched a center to focus on building people-pleasing AIs.
The Center for Human-Compatible Artificial Intelligence, launched this week with $5.5m in funding from the Open Philanthropy Project, is lead by computer science professor and artificial intelligence pioneer Stuart Russell. Heâ€™s quick to dispel any â€œunreasonable and melodramaticâ€ comparisons to the threats posed in science fiction.
â€œThe risk doesnâ€™t come from machines suddenly developing spontaneous malevolent consciousness,â€ he said. â€œItâ€™s important that weâ€™re not trying to prevent that from happening because thereâ€™s absolutely no understanding of consciousness whatsoever.â€
Russell is well known in the artificial intelligence community and in 2015 penned an open letter calling for researchers to look beyond the goal of simply making AI more capable and powerful to think about maximizing its social benefit. The letter has been signed by more than 8,000 scientists and entrepreneurs including physicist Stephen Hawking, entrepreneur Elon Musk and Apple co-founder Steve Wozniak.
â€œThe potential benefits [of AI research] are huge, since everything that civilization has to offer is a product of human intelligence; we cannot predict what we might achieve when this intelligence is magnified by the tools AI may provide, but the eradication of disease and poverty are not unfathomable,â€ the letter reads.
â€œBecause of the great potential of AI, it is important to research how to reap its benefits while avoiding potential pitfalls.â€
Itâ€™s precisely this thinking that underpins the new center.
Up until now, AI has primarily been applied to very limited contexts such as playing Chess or Go or recognizing objects in images, where there isnâ€™t much scope for the system to do much damage. As they start to make decisions on our behalf within the real world, the stakes are much higher.
â€œAs soon as you put things in the real world, with self-driving cars, digital assistants â€¦ as soon as they buy things on your behalf, turn down appointments, then they have to align with human values,â€ Russell said.
He uses autonomous vehicles to illustrate the type of problem the center will try to solve. Someone building a self-driving car might instruct it never to go through a red light, but the machine might then hack into the traffic light control system so that all of the lights are changed to green. In this case the car would be obeying orders but in a way that humans didnâ€™t expect or intend. Similarly, an artificially intelligent hedge fund designed to maximize the value of its portfolio could be incentivized to short consumer stocks, buy long on defence stocks and then start a war â€“ as suggested by Elon Musk in Werner Herzogâ€™s latest documentary.
â€œEven when you think youâ€™ve put fences around what an AI system can do it will tend to find loopholes just as we do with our tax laws. You want an AI system that isnâ€™t motivated to find loopholes,â€ Russell said.
â€œThe problem isnâ€™t consciousness, but competence. You make machines that are incredibly competent at achieving objectives and they will cause accidents in trying to achieve those objectives.â€
To address this, Russell and his colleagues at the center propose making AI systems that observe human behavior and try to work out what the humanâ€™s objective is, then behave accordingly and learn from mistakes. So instead of trying to give the machine a long list of rules to follow, the machine is told that its main objective is to do what the human wants them to do.
It sounds simple, but itâ€™s not how engineers have been building systems for the past 50 years.
But if AI systems can be designed to learn from humans in this way, it should ensure that they remain under human control even when they develop capabilities that exceed our own.
In addition to watching humans directly using cameras and other sensors, robots can learn about us by reading history books, legal documents, novels, newspaper stories as well as by watching videos and movies. From this they can start to build up an understanding of human values.
It wonâ€™t be easy for machines. â€œPeople are irrational, inconsistent, weak-willed, computationally limited, heterogenous and sometimes downright evil,â€ Russell said.
â€œSome are vegetarians and some really like a nice juicy steak. And the fact that we donâ€™t behave anything close to perfectly is a serious difficulty.â€
By Olivia Solon